Skip to main content

Image Preprocessing

Module: src-tauri/src/pipeline/preprocess.rs Function: preprocess_floor_plan(input_path: &str) -> Result<String, String>

The preprocessing stage prepares raw floor plan images for downstream parsing. It runs three sequential transformations: rotation correction, border cropping, and size limiting.

Processing Steps​

Step 1: Resize (Size Limiting)​

Limits the longest edge to 2048 pixels using Lanczos3 interpolation. This runs first to cap memory usage before heavier processing.

fn resize_if_large(img: DynamicImage, max_size: u32) -> DynamicImage
  • If the longest edge is already <= 2048px, the image passes through unchanged
  • Otherwise, scales down proportionally so the longest edge equals 2048px
  • Uses FilterType::Lanczos3 for high-quality downscaling

Example: A 3000x2000 image becomes 2048x1365.

Step 2: Rotation Correction​

Corrects slight tilts (within +/-5 degrees) that are common in photographed or scanned floor plans.

Algorithm:

  1. Canny edge detection -- runs on the grayscale image with thresholds (50.0, 150.0)
  2. Gradient angle sampling -- for each edge pixel, computes a Sobel-like gradient direction
  3. Angle normalization -- maps angles to near-horizontal/vertical range
  4. Filtering -- keeps only angles with |angle| < 5.0 degrees
  5. Median angle -- computes the median of all collected angles
  6. Rotation transform -- if median angle >= 0.5 degrees, rotates the image using bilinear interpolation with white fill
fn correct_rotation(img: DynamicImage, gray: &GrayImage) -> DynamicImage

The rotation uses imageproc::geometric_transformations::rotate_about_center.

Step 3: Border Cropping​

Removes white borders around the floor plan content area.

Algorithm:

  1. Threshold -- scan all pixels, find those with intensity < 240 (non-white)
  2. Bounding box -- compute the min/max x/y of all non-white pixels
  3. Padding -- add 20px padding around the content bounding box
  4. Clamp -- ensure the crop region stays within image bounds
  5. Safety check -- skip cropping if the resulting region is smaller than 100x100px
fn crop_content(img: DynamicImage) -> DynamicImage

Input​

Raw floor plan image in any common format (JPG, PNG):

data/uploads/project_abc/floorplan.jpg   (3000x2000, 2.5MB)

Output​

A preprocessed temporary file, also copied to the pipeline directory:

/tmp/planova_processed_{uuid}.jpg        (1800x1200, ~800KB)
data/pipeline/{project_id}/preprocessed.jpg

Implementation Notes​

  • The grayscale conversion (to_luma8()) is used for edge detection and content detection, but the actual transforms are applied to the full RGBA image
  • The rotation correction skips processing if the median angle is less than 0.5 degrees (sub-pixel alignment)
  • The crop threshold of 240 (out of 255) handles light gray borders but preserves white space within the floor plan that may be part of room interiors
  • The 20px crop padding ensures wall lines at the edges of the content area are not clipped